Source-linked AI summary
Monocular Mesh Recovery and Body Measurement of Female Saanen Goats
Bo Jin, Shichao Zhao, Jin Lyu, Bin Zhang, Tao Yu, Liang An, Yebin Liu, Meili Wang
TL;DR
Saanen goat body measurement is important for precision livestock management, but existing methods lack authentic goat-specific 3D data and suitable models. The paper builds an eight-view RGBD dataset and a SaanenGoat parametric model from high-fidelity scans, enabling single-view reconstruction and automated measurement of six body dimensions. The method outperforms generic baselines in reconstruction and measurement accuracy, while remaining limited in cross-breed generalization and occluded mammary-region reconstruction.
Problem
Precision Saanen goat management needs accurate 3D body measurement, but existing approaches lack high-quality goat data and models representing structures such as the udder.
Method
The paper captures synchronized eight-view RGBD videos, fuses them with multi-view DynamicFusion, and fits a 41-joint SaanenGoat model using scans from 48 goats.
Results
The method enables single-view RGBD mesh recovery and automated measurement of six body dimensions, reducing reconstruction errors by up to 77.7% on In-Shape and 66.5% on Out-Shape versus SMAL.
Takeaways & Limitations
The framework provides a specialized foundation for non-invasive biometric measurement and 3D livestock analysis in precision farming.
Takeaways & Limitations
Generalization is constrained to the target species, and reconstruction accuracy is suboptimal for occluded mammary regions.
Abstract
from arXiv · showhide
The lactation performance of Saanen dairy goats, renowned for their high milk yield, is intrinsically linked to their body size, making accurate 3D body measurement essential for assessing milk production potential, yet existing reconstruction methods lack goat-specific authentic 3D data. To address this limitation, we establish the FemaleSaanenGoat dataset containing synchronized eight-view RGBD videos of 55 female Saanen goats (6-18 months). Using multi-view DynamicFusion, we fuse noisy, non-rigid point cloud sequences into high-fidelity 3D scans, overcoming challenges from irregular surfaces and rapid movement. Based on these scans, we develop SaanenGoat, a parametric 3D shape model specifically designed for female Saanen goats. This model features a refined template with 41 skeletal joints and enhanced udder representation, registered with our scan data. A comprehensive shape space constructed from 48 goats enables precise representation of diverse individual variations. With the help of SaanenGoat model, we get high-precision 3D reconstruction from single-view RGBD input, and achieve automated measurement of six critical body dimensions: body length, height, chest width, chest girth, hip width, and hip height. Experimental results demonstrate the superior accuracy of our method in both 3D reconstruction and body measurement, presenting a novel paradigm for large-scale 3D vision applications in precision livestock farming.
Introduction
Saanen goats’ body characteristics correlate with milk production, but precision 3D management is limited by scarce high-quality data and inadequate goat-specific models. This work introduces a dynamic RGBD dataset and SaanenGoat model for reconstruction and body measurement.
- Saanen goats’ morphometric characteristics are strongly correlated with milk production, motivating precise monitoring and phenotyping.
- 3D vision deployment is hindered by difficult on-farm data acquisition and models that inadequately represent structures such as the udder.
- FemaleSaanenGoat provides synchronized eight-view RGBD videos of 55 goats aged 6–18 months in diverse natural postures.Each sequence lasts 30 seconds at 30 FPS.
- Multi-view DynamicFusion converts noisy point-cloud sequences into approximately 3,200 high-fidelity 3D scans, with 32 manually annotated anatomical keypoints per scan.Approximately 2,700 scans are used for training and 500 for testing.
- SaanenGoat is a real-data-driven parametric model with anatomically relevant mammary-gland representation for single-view RGBD reconstruction and six-dimensional body measurement.
- 3D reconstruction errors decrease by up to 77.7% on In-Shape and 66.5% on Out-Shape versus SMAL, while measurement MAE reaches 1.90 versus 4.89 for SMAL and 3.48 for SMAL+.
Related Works
Prior quadruped models are constrained by limited or mismatched shape spaces, while monocular animal recovery methods remain poorly suited to rare agricultural animals. Species-specific modeling addresses these limitations but has focused mainly on other animals.
- SMAL was built from 41 toy-animal scans and has limited shape space, preventing adequate representation of out-of-domain species such as goats.
- SMAL extensions add animal-specific displacements or dog scale parameters but retain its bone structure and mean shape.
- Species-specific parametric models include systems for dogs, horses, and other animals, using breed information, anatomical skeletons, or muscle-based deformations.
- Monocular recovery methods span model-free articulated or implicit approaches and model-based SMAL methods, but rare agricultural animals remain challenging.
3D Body Measurement for Agriculture
Agricultural body measurement is moving from stressful, subjective manual procedures toward computer-vision and 3D-imaging workflows. The paper’s pipeline uses multi-view RGBD capture and dynamic fusion to produce goat scans for model-based measurement.
- Manual livestock measurement is labor-intensive, time-consuming, subjective, stressful, and prone to inconsistency.
- Prior computer-vision systems use multi-view depth or RGB keypoints projected onto point clouds to calculate livestock body measurements.
- The acquisition pipeline combines one top-view and seven surrounding RGBD cameras to capture synchronized goat data.
- Multi-view DynamicFusion generates geometrically accurate goat models, which are registered to an initialization template and normalized to T-pose for shape-space construction.
- The dataset includes scan results spanning goats of different ages and body types.
Method
The method builds goat-specific 3D data and a SaanenGoat parametric model, then fits scans and recovers shape variation using pose normalization and PCA-based blend shapes. Single-view recovery adds depth, contour-normal, and biomechanical constraints while preserving anatomical asymmetries.
- Data acquisition and reconstruction: Eight synchronized RGBD cameras capture goat point clouds for high-quality 3D reconstruction.The setup uses one top view and seven surrounding views; DynamicFusion is modified for eight-view deforming surfaces.
- Parametric model construction: The SaanenGoat template expands to 13,815 vertices and 41 skeletal joints, with a kinematic tree defining joint-parent relationships.The template is constructed specifically from Saanen dairy goat scans rather than relying on the generic SMAL model.
- Registration and fitting: Pose fitting uses 32 anatomical landmarks, rigid SVD alignment, recursive kinematics, linear blend skinning, and losses for scale, landmarks, correspondence, collisions, and pose regularization.The landmark alignment establishes coarse correspondence before fine-scale deformation analysis.
- Registration and fitting: Shape fitting refines the pose-aligned template with vertex displacements and partitions optimization into torso, head, and limb regions.The regional partition addresses inconsistent vertex density and captures anatomical morphology with high fidelity.
- Shape space: The shape space normalizes 48 fitted goats to a T-pose, computes deviations from a standard template, and applies PCA to obtain orthogonal shape bases.Figure 5 visualizes the first four principal components at ±2 standard deviations.
- Shape space: The model generates varied goat body types by combining pose and low-dimensional shape parameters, while omitting pose blend shapes because they produced negligible improvements and preserving anatomical asymmetries.No vertex symmetry enforcement is applied because asymmetries occur in structures including the lungs and mammary glands.
- Monocular mesh recovery: Monocular recovery incorporates depth and contour-normal constraints plus biomechanical regularization to resolve spatial ambiguity and prevent physiologically implausible distortions.These additions replace SMAL’s generic model and its reliance only on 2D keypoints and silhouettes.
Experiments
Experiments evaluate SaanenGoat against SMAL, SMAL+, and PointStack-based measurements using scanned goats and single-view RGBD inputs. The results show stronger reconstruction and body-measurement accuracy, including improved performance on unseen goats and substantial monocular error reductions.
- 3D reconstruction: The In-Shape and Out-Shape evaluations compare SMAL, SMAL+, and SaanenGoat using reconstruction errors on standing, walking, head-turning, backward-looking, and head-lowering poses.The In-Shape test set contains 300 frames from five pose categories, while the Out-Shape set contains 350 frames from seven goats excluded from shape-space construction.
- 3D reconstruction: SaanenGoat achieves the lowest fitting errors because its learned shape space captures authentic Saanen goat body structure more accurately than SMAL and SMAL+.Quantitative and qualitative results also show accurate representation across various postures and generalization to unseen Saanen goats.
- Body measurement: The measurement system estimates six dimensions: body length, body height, hip height, chest width, hip width, and chest girth.Keypoint pairs provide Euclidean distances for body length, chest width, and hip width; heights use the ground plane, and chest girth uses the model’s anatomical keypoints.
- Body measurement: Compared with PointStack, the model reduces MAE for all six body parts, with the largest chest-girth improvement reaching 45.6%.For MAPE, it outperforms PointStack on five of six metrics, including 46.7% lower chest-girth error and 26.0% lower hip-width error.
- Monocular mesh recovery: In monocular recovery, SAANEN achieves 70.8% and 65.9% lower errors than SMAL and SMAL+, respectively.For T-pose reconstructed meshes, average MAE is 20.6 mm versus 62.8 mm for SMAL and 41.0 mm for SMAL+.
Conclusion
The paper presents an authentic eight-view RGBD dataset and a specialized SaanenGoat model for 3D reconstruction and non-invasive biometric measurement. It concludes that the framework supports livestock analysis, while remaining limited by breed specificity and occluded mammary regions.
- Conclusion: FemaleSaanenGoat contains 55 female Saanen goats with diverse natural postures captured in synchronized eight-view RGBD videos.Multi-view Dynamic Fusion produces high-fidelity 3D reconstructions used to develop the specialized parametric model.
- Conclusion: SaanenGoat represents goat morphology with interpretable pose and shape parameters and automatically extracts six biometric measurements from anatomical landmarks.The reported measurements include body length, withers height, shoulder breadth, thoracic circumference, pelvic width, and hip height.
- Limitations: Generalization to other breeds is limited, and reconstruction accuracy remains suboptimal for occluded mammary regions.Proposed future work includes occlusion-aware reconstruction, multi-breed datasets, and multimodal sensing for varying wool thickness.
- Conclusion: The framework provides a transferable foundation for 3D vision-based livestock analysis and AI-driven agricultural applications.This conclusion is stated within the scope of precision farming and livestock industry digitalization.